void operator()( tbb::blocked_range<int> rng ){
double samples_per_image = 1.0*n_samples / lab_images.count();
// Collect some statistics on mean and variance of each feature
for( int i=rng.begin(), n_features=rng.begin()*samples_per_image; i<rng.end(); i++ ){
Image< float > feature_response = feature->evaluate( lab_images[i], names[i] );
for( int j=0; j<feature_response.height(); j++ )
for( int i=0; i<feature_response.width(); i++ ){
VectorXd x( feature_size );
for( int k=0; k<feature_size; k++ )
x[k] = feature_response(i,j,k);
count += 1;
VectorXd delta = x - mean;
mean += delta / count;
covariance += delta * (x-mean).transpose();
}
for(;n_features < (i+1)*samples_per_image; n_features++){
int x = random() % feature_response.width();
int y = random() % feature_response.height();
for( int i=0; i<feature_size; i++ )
features.append( feature_response( x, y, i ) );
}
}
}
void computeFeatures( QVector<float> & features, VectorXd & mean, MatrixXd & covariance, const QSharedPointer<Feature> & feature, const QVector< Image< float > >& lab_images, const QVector< QString >& names, int n_samples ){
int feature_size = feature->size();
int n_features = 0;
double samples_per_image = 1.0*n_samples / lab_images.count();
// Do propper whitening: http://en.wikipedia.org/wiki/White_noise#Whitening_a_random_vector
// Collect some statistics on mean and variance of each feature
mean = VectorXd::Zero( feature_size );
covariance = MatrixXd::Zero( feature_size, feature_size );
double count = 0;
for( int i=0; i<lab_images.count(); i++ ){
Image< float > feature_response = feature->evaluate( lab_images[i], names[i] );
for( int j=0; j<feature_response.height(); j++ )
for( int i=0; i<feature_response.width(); i++ ){
VectorXd x( feature_size );
for( int k=0; k<feature_size; k++ )
x[k] = feature_response(i,j,k);
count += 1;
VectorXd delta = x - mean;
mean += delta / count;
covariance += delta * (x-mean).transpose();
}
for(;n_features < (i+1)*samples_per_image; n_features++){
int x = random() % feature_response.width();
int y = random() % feature_response.height();
for( int i=0; i<feature_size; i++ )
features.append( feature_response( x, y, i ) );
}
}
// Compute the final covariance
covariance = covariance / count;
}
Image< short > Texton::textonize(const Image< float >& lab_image, const QString & name ) const{
Image< float > feature_response = feature_->evaluate( lab_image, name );
// Whitening (zero mean, 1 stddev)
for( int j=0; j<feature_response.height(); j++ )
for( int i=0; i<feature_response.width(); i++ ){
VectorXd x( feature_response.depth() );
for( int k=0; k<feature_response.depth(); k++ )
x[k] = feature_response(i,j,k);
x = transformation_*(x - mean_);
for( int k=0; k<feature_response.depth(); k++ )
feature_response(i,j,k) = x[k];
}
return kmeans_.evaluate( feature_response );
}
// [ref] ${VXL_HOME}/contrib/mul/fhs/tools/find_matches.cxx
int fhs_find_matches_example(int argc, char *argv[])
{
vul_arg<vcl_string> image1_path("-i1", "Input image 1");
vul_arg<vcl_string> image2_path("-i2", "Input image 2");
vul_arg<vcl_string> output_image1_path("-o1", "Output image 1", "output1.png");
vul_arg<vcl_string> output_image2_path("-o2", "Output image 1", "output2.png");
vul_arg<unsigned> level("-L", "Image pyramid level to work on", 2);
vul_arg<unsigned> nc("-n", "Number of points to select", 10);
vul_arg<unsigned> w("-w", "Half width of filters", 7);
vul_arg<double> f("-f", "Relative strength of intensity vs shape", 0.1);
vul_arg_parse(argc, argv);
if (image1_path() == "" || image2_path() == "")
{
local::print_usage();
return 0;
}
// ============================================
// Attempt to load in images
// ============================================
vimt_image_2d_of<vxl_byte> image1, image2;
image1.image() = vil_load(image1_path().c_str());
if (image1.image().size() == 0)
{
vcl_cerr << "Unable to read in image from " << image1_path() << vcl_endl;
return 1;
}
image2.image() = vil_load(image2_path().c_str());
if (image2.image().size() == 0)
{
vcl_cerr <<" Unable to read in image from " << image2_path() << vcl_endl;
return 1;
}
// ============================================
// Build image pyramids and select chosen level
// ============================================
vimt_gaussian_pyramid_builder_2d<vxl_byte> pyr_builder;
vimt_image_pyramid image_pyr1, image_pyr2;
pyr_builder.build(image_pyr1, image1);
pyr_builder.build(image_pyr2, image2);
const vimt_image_2d_of<vxl_byte> &image1_L = static_cast<const vimt_image_2d_of<vxl_byte> &>(image_pyr1(level()));
const vimt_image_2d_of<vxl_byte> &image2_L = static_cast<const vimt_image_2d_of<vxl_byte> &>(image_pyr2(level()));
// ====================================================
// Apply corner operator to image1_L and select corners
// ====================================================
vimt_image_2d_of<float> corner_im;
corner_im.set_world2im(image1_L.world2im());
vil_corners(image1_L.image(), corner_im.image());
vcl_vector<unsigned> pi, pj;
float threshold = 4.0f;
vil_find_peaks_3x3(pi, pj, corner_im.image(), threshold);
// Evaluate corner strength at each point (pi[i], pj[i])
unsigned n = pi.size();
vcl_vector<float> corner_str(n);
for (unsigned i = 0 ; i < n; ++i)
corner_str[i] = corner_im.image()(pi[i], pj[i]);
// Sort and generate a list of image points and equivalent world points
vcl_vector<unsigned> index;
mbl_index_sort(corner_str, index);
unsigned n_c = vcl_min(nc(),n);
vcl_vector<vgl_point_2d<int> > im_pts(n_c);
vcl_vector<vgl_point_2d<double> > w_pts(n_c);
vimt_transform_2d im2w = image1_L.world2im().inverse();
for (unsigned i = 0; i < n_c; ++i)
{
im_pts[i] = vgl_point_2d<int>(pi[index[n - 1 - i]], pj[index[n - 1 - i]]);
w_pts[i] = im2w(pi[index[n - 1 - i]], pj[index[n - 1 - i]]);
}
// ========================================================
// Extract patches around each selected point and normalise
// ========================================================
vcl_vector<vil_image_view<float> > patch(n_c);
vcl_vector<vgl_point_2d<double> > patch_ref(n_c); // Reference point
int ni = image1.image().ni();
int nj = image1.image().nj();
for (unsigned i = 0; i < n_c; ++i)
{
// Select region around point, allowing for image edges.
int ilo = vcl_max(0, int(im_pts[i].x() - w()));
int ihi = vcl_min(ni - 1, int(im_pts[i].x() + w()));
int jlo = vcl_max(0, int(im_pts[i].y() - w()));
int jhi = vcl_min(nj - 1, int(im_pts[i].y() + w()));
// Compute position of reference point relative to corner
int kx = im_pts[i].x() - ilo;
int ky = im_pts[i].y() - jlo;
patch_ref[i] = vgl_point_2d<double>(kx, ky);
vil_convert_cast(vil_crop(image1_L.image(), ilo, 1 + ihi - ilo, jlo, 1 + jhi - jlo), patch[i]);
vil_math_normalise(patch[i]);
}
// Construct tree structure for points
vcl_vector<vcl_pair<int, int> > pairs;
mbl_minimum_spanning_tree(w_pts, pairs);
assert(pairs.size() == n_c - 1);
// Draw tree into image for display purposes
local::draw_tree(image1.image(), w_pts, pairs);
if (vil_save(image1.image(), output_image1_path().c_str()))
{
vcl_cout << "Saved output image 1 to " << output_image1_path() << vcl_endl;
}
// =================================================
// Construct the arc model from the points and pairs
// =================================================
vcl_vector<fhs_arc> arcs(n_c - 1);
int root_node = pairs[0].first;
for (unsigned i = 0; i < pairs.size(); ++i)
{
int i1 = pairs[i].first;
int i2 = pairs[i].second;
vgl_vector_2d<double> dp = w_pts[i2] - w_pts[i1];
double sd_x = vcl_max(0.1 * ni, 0.2 * dp.length());
double sd_y = vcl_max(0.1 * nj, 0.2 * dp.length());
arcs[i] = fhs_arc(i1, i2, dp.x(), dp.y(), sd_x * sd_x, sd_y * sd_y);
}
// =================================================
// Apply filters to image2 (initially to whole image)
// =================================================
vcl_vector<vimt_image_2d_of<float> > feature_response(n_c);
for (unsigned i = 0; i < n_c; ++i)
{
// Apply to whole image in first instance
// Ideally would crop a region around expected position
vimt_normalised_correlation_2d(image2_L, feature_response[i], patch[i], patch_ref[i], float());
// Need good values to be small, not large, so apply -ve factor
vil_math_scale_values(feature_response[i].image(), -f());
}
// ======================================================
// Use fhs_searcher to locate equivalent points on image2
// ======================================================
fhs_searcher searcher;
searcher.set_tree(arcs, root_node);
searcher.search(feature_response);
vcl_vector<vgl_point_2d<double> > pts2;
searcher.best_points(pts2);
// Draw tree into image for display purposes
local::draw_tree(image2.image(), pts2, pairs);
if (vil_save(image2.image(), output_image2_path().c_str()))
{
vcl_cout << "Saved output image 2 to " << output_image2_path() << vcl_endl;
}
return 0;
}
